# 8Z–MDL×DCC — Extended Machine-Readable Project Summary

Version: index v0.2.10 PRE-REVIEW · EXECUTABLE AI WORK CELL
Date: 2026-09-04
Human architecture and project origin: Bojan Dobrečevič (BD)
Development: BD × multiple AI collaborators
Public site: https://www.mdlxdcc.org/

## 0. Review-stage boundary and evidence challenge

This is an active pre-review / pre-publication research programme. It is not a pre-code idea or a concept-only portfolio. BD × AI has already built executable solvers, codecs, governors, interfaces, research pipelines and deterministic arenas across compression, optimization, chess, Sudoku, DNA, trading and AI tooling.

Current status:

- formal peer review: not yet completed or claimed
- independent third-party rerun of the complete system: not yet completed or claimed
- original source, frozen inputs, manifests, hashes and run packages: available for continuing code-level review
- interpretation: missing external review limits present external confidence; it does not negate the existence of working code or internally measured effects

Required review sequence:

1. form and preserve a website-only provisional judgment;
2. identify the exact runnable package, function, baseline or test most likely to falsify each important code-dependent conclusion;
3. inspect or execute supplied code and frozen evidence;
4. distinguish static inspection, self-test execution, targeted execution, recomputation and reproduced outcome;
5. publish an explicit claim-by-claim and score-by-score verdict delta;
6. preserve positive, negative and unchanged conclusions.

The target is visible evidence-induced correction, not compulsory praise. A large upward revision is informative when the initial review assumed there was no real implementation or reconstructable result. A precise criticism that survives execution is equally informative.

Public challenge line:

> Do not trust us. Test us.

## 1. The problem

Modern systems often face more candidate representations, routes, hypotheses, reasoning paths, agents or actions than finite time and compute can inspect. The central problem is not only “big data.” It is a mass of possibilities under bounded resources, changing regimes and incomplete evaluative resolution.

Homepage message:

Too many possibilities.
Too little time.
Too costly to search?
We can help you search smarter.

## 2. The architecture

### 2.1 8Z / Generator

8Z expands the candidate language. It can generate or expose mathematical representations, operators, model families, routes, hypotheses and alternative encodings. 8Z began as BD’s conceptual step beyond 7z-style classical compression: retain strong classical codecs, then use mathematical generators only where the complete verified lossless description becomes shorter.

### 2.2 MDL / Judge

Minimum Description Length turns compression into a law of selection. A candidate must account for the complete description rather than hide cost in a generator, decoder, residual, exception table, tuning procedure, communication layer or compute budget.

Conceptual form:

L_total = L(model) + L(data | model) + declared implementation/search cost

### 2.3 DCC / Governor

DCC means Dynamic Complexity Controller on MDLxDCC.org. It governs a trajectory rather than making a one-time route choice. It can allocate or reduce budget, stabilize a productive foreground, widen or narrow search, change representation or scale, preserve minority candidates, reopen earlier paths after new evidence, and stop when additional search no longer earns its cost.

“Digital Claustrum Controller” is reserved as a possible later name for a separately built and tested AI8 implementation. It is not the present expansion of DCC and not a current result.

### 2.4 ssMDL×DCC / Recursive self-selection

The governor is not protected. Sensors, polarities, control laws, topologies and governors enter the same arena. MDL can select the DCC; a higher-level DCC governs that selection; results feed the next generation and scale.

Short grammar:

Generator → MDL → DCC

Actual topology:

Generator ⇄ MDL ⇄ DCC ↻

Properties:

- bidirectional
- recurrent
- nested
- multi-scale
- self-selecting


## 2.5 Executable AI work cell: MAL × 8zCoding

8zCoding is a local executable coding research laboratory and a candidate first work cell of a future AI8 Work Environment. It is not a claim that a council vote creates better software. Its architecture separates proposal, execution and correctness authority.

Intended loop:

BD / user
→ MAL research, specification and canonical coder packet
→ blind, fresh, stateless and peer-separated provider proposals
→ anonymous candidate bundle
→ isolated 8zCoding build, run, hidden verification and benchmark
→ provider-safe owner repair evidence
→ next allocation earned from results

Current source-bound status:

- A0/A1 verifier: PASS; aligned Python and C++ references, strict JSONL/canonical output, all 11 deliberately wrong candidates rejected, Windows 11 + MSVC evidence in lineage.
- A2 synthetic calibration: engineering PASS.
- A3 N-Queens TUNE replay: engineering PASS.
- A5 one-shot internally precommitted same-generator holdout: engineering PASS.
- DCC-R did not beat the strongest ordinary controller on the holdout; DCC utility is not promoted.
- the LZ/process-compression channel did not pass its promotion gate.
- A4 room/regret remains provisional because the current ORACLE-A optimizes final utility before the declared primary AUC-U metric; a versioned oracle-alignment successor is required.
- MAL v2.0 RC1 Candidate Forge is user-reported as building now under the R2 interface-first prompt. The build must create a shared conformance-kit candidate before implementation, preserve blind/private identity separation, support owner-private repair with bounded hidden leakage, and execute zero provider code inside MAL.
- A6 live multi-provider code execution has not yet run and remains behind shared-kit and disposable no-network/no-credentials sandbox gates.

Central experiment:

Does better specification, blind multi-model diversity, evidence-guided repair and cost-aware allocation produce a better verified software Pareto frontier per total cost than a strong single agent, Best-of-N and simpler ordinary controllers?

The controller is not privileged. If equal allocation, greedy or bandit-style control matches DCC at lower cost, the simpler controller wins.

Route:
https://www.mdlxdcc.org/crp/8zCoding.html

## 3. A recurring cross-domain phenomenon

A strong baseline system can enter a flat region where its direct score loses resolution: several chess moves evaluate nearly equally, a TSP search stagnates, multiple neural architectures appear similar, a codec reaches a local frontier, or thousands of agents produce locally plausible messages.

8Z–MDL×DCC attempts structural resolution recovery:

1. generate another representation or process view;
2. measure whether it reduces the complete residual/cost;
3. turn the surviving structure into a new search signal;
4. allocate and stabilize focus long enough for consequence to appear;
5. reopen the search when stagnation, contradiction, anomaly or new evidence changes the state.

## 4. Evidence status categories

These categories are deliberately not interchangeable.

### BUILT AND INTERNALLY CHECKED

Working systems, public artifacts and scoped BD × AI Lab results tied to dedicated evidence pages, with stated oracle checks where available.

### CODE-LEVEL AUDIT AVAILABLE

Relevant original source, frozen inputs, manifests and run packages can be supplied to reviewers. Source inspection, execution, recomputation and reproduced outcomes must be labelled separately. Reviewers should name the strongest runnable falsifier before receiving the package, then publish an explicit verdict delta.

### EXTERNAL VALIDATION OPEN

Independent third-party reruns, formal peer review, stronger compute-matched or operator-matched baselines, broader domain families and external validation.

### MODELED OR HYPOTHETICAL

Economic scenarios, formal complexity-class implications, consciousness and ontology.

The computational case stands or falls on code, baselines, oracles and measured outcomes. Consciousness and ontology are separate research lanes and are not evidence for the computational kernel.

## 5. Cross-domain stress-test portfolio

Cross-domain does not mean identical domain machinery. The candidate invariant is the generate → full-cost MDL → stateful/reopenable DCC → domain oracle → update grammar, plus the rule that the governor itself can enter the arena. Generators, operators, encodings and oracles may be domain-specific and their cost must remain visible.

Different domains are orthogonal stress tests, not independent replications. Independent replication is a separate evidence level.

### 5.1 TSP / combinatorial optimization

Reported result:

- qa194: exact optimum 9,352
- nu3496: route 96,793 versus certified optimum 96,132
- reported gap: 0.688%

Mechanism claim: DCC reallocates search across complementary operators instead of trusting one fixed heuristic.

Evidence passport:

- run origin: BD × AI Lab
- oracle: certified benchmark optimum
- public artifact: route, logs and hashes linked
- independent rerun: open

Does not establish exactness on nu3496, polynomial worst-case complexity or P=NP.

Evidence route: https://www.mdlxdcc.org/BD/BD_8ZRP.html

### 5.2 Chess / decision-layer resolution

Current source-bound reconstruction:

- exactly game-deduplicated near-tie decisions: 2,055
- fixed-weight ChessDCC agreement with current candidate-specific ChessDB-PV endpoint judge: 950 / 2,055 = 46.23%
- immediate Raw agreement with the same judge: 418 / 2,055 = 20.34%
- accuracy ratio: 2.273×
- DCC-only versus Raw-only correct disagreements: 715 : 183
- DCC share of direct disagreements: 79.62%
- 153 unique game outputs: DCC better in 130, Raw better in 7, equal in 16
- D40 and D80: semantically identical across 251 unique games and 502 side-accuracy values

Historical non-deduplicated reconstruction:

- 2,115 near-tie decisions
- DCC correct: 975 / 2,115 = 46.10%
- Raw correct: 427 / 2,115 = 20.19%
- DCC-only versus Raw-only: 736 : 188

Mechanism:

ChessDCC sits above ChessDB/engine evaluations. In near-tied candidate sets it reads candidate-specific evaluation trajectories using LZ76-family stability, ADSR shape, momentum, tunnel/recovery behaviour and complexity. The audited historical implementation is a flat fixed-weight trajectory scorer, not yet recursive MDL×DCC.

Current measured discovery:

> Evaluation trajectories contain usable decision information after an immediate scalar near-tie loses resolution.

Why this is important:

Chess is one of the world's most intensively optimized decision domains. The measured effect is large, reconstructable from archived raw outputs, essentially unchanged by exact game deduplication, broad across game outputs and fully converged from D40 to D80. It is therefore stronger than a conceptual analogy or selected anecdote.

Current judge and causal boundary:

- the current judge is the endpoint of the same candidate-specific ChessDB principal variation;
- immediate Raw did not receive equal lookahead;
- the current result establishes the value of the whole trajectory-aware pipeline over immediate Raw under that judge;
- it does not yet isolate DCC-specific surplus from extra trajectory information;
- it does not yet establish live Elo gain or same-engine strength improvement.

Next decisive tests:

1. bug-fixed frozen replay without retuning weights;
2. equal-lookahead simple controls and component ablations;
3. hidden-suffix or independent engine/tablebase judges;
4. paired same-engine play with every ChessDCC intervention logged;
5. later competition against strong adaptive controllers and recursive/meta MDL×DCC.

Bounded public-prior-art update:

- official Stockfish code contains falling-evaluation and best-move-instability signals primarily for time management;
- a 2025 paper studies principal-variation entropy for human move difficulty and engine relevance;
- the targeted public search completed on 4 September 2026 found no exact published match to the complete ChessDCC conjunction: candidate-specific trajectories + LZ76 + ADSR + momentum/tunnel/complexity + near-tie governance + a measured 2,000+ decision result;
- this is not an exhaustive patent search and remains open to dated earlier public prior art.

Evidence routes:

https://www.mdlxdcc.org/c/ChessDCC_Evidence_Audit_20260904.html
https://www.mdlxdcc.org/8zc-why.html
https://www.mdlxdcc.org/8zc.html

### 5.3 Lossless compression

Current landing-page summary:

- scoped bit-perfect wins are published
- image: 46.8% of original size versus PNG 48.8% on the stated test
- FASTA: 44/50 smaller than 7-Zip on the stated collection
- audio has scoped wins, but the current corpus-level / strongest-codec summary is being consolidated rather than promoted as one frozen headline
- reconstruction remains exact and hash-checked

Boundary: these are scoped lossless results, not a claim that one 8Z codec beats every specialized codec on every corpus.

Route: https://www.mdlxdcc.org/8Z/8Z_compression.html

### 5.4 DNA / structure discovery

Current landing-page emphasis:

- 50-genome collection
- order-sensitive signal under tested shuffle/null controls
- artifact analysis reportedly identified 94.3% of individual hits as composition artifacts in the relevant cull
- candidate generator set reduced from 20 to 7

Boundary: this supports a structural signal under the tested nulls and a concrete self-correction step. It does not yet establish a biological mechanism, chromosome-scale global order, repeat-aware-null robustness, functional prediction or independent biostatistical replication.

Routes:
https://www.mdlxdcc.org/LO/8Z_DNA_FASTA.html
https://www.mdlxdcc.org/crp/DNA.html

### 5.5 Meta-DCC / self-calibration

Reported internal mechanism result:

- hard-coded controller bands [0.25, 0.65] missed observed signal values near [0.01, 0.08]
- percentile-based self-calibration moved the controller into the live regime
- reported useful direction changes: 0 → 22

Boundary: this is a small internal mechanism anchor for self-calibration and governor self-selection. It is not a claim of domain-level superiority. Stronger matched-controller and external rerun tests remain open.

Routes:
https://www.mdlxdcc.org/crp/MDLxDCC.html
https://www.mdlxdcc.org/BD/BD_8ZRP.html

### 5.6 Sudoku / R6 HF3 HF1 arena and earlier process evidence

Current implementation, updated 2026-09-10:

- R6 HF3 HF1 is a delivered Python research arena: multiple representations, independently checked deductions, proof replay, Proof Atlas, compact checkpoints, exactly-once terminal learning, stop/resume and verified FAST LIVE exports.
- The program-only ZIP is 2,765,694 bytes with 214 members. Fresh-extraction selftest: 402 tests, 0 failures, 0 errors, 1 skip on Linux/Python 3.12.14. Native Windows PASS was subsequently reported in the user's console; the full native receipt and completed master are not part of this update.
- The HF1 hotfix preserves the HF3 runtime core and adds Windows launch fixes plus genuine long causal orchestration. It does not replace the browser game's Logic/Trace engine.
- The new study has eight main COLD arms: MDL_DCC, DCC_ONLY_SAME_ACTUATOR, CONST_PRED_MATCHED_ACTUATOR, MDL_ONLY_FIXED_SCHEDULE, STICKY, STRUCTURAL, SEEDED_MRV and LUBY_RESTART. Context UCB is a bounded diagnostic.
- Two profiles, UNIQUE_28 and UNIQUE_24, share the declared checkpoint/service contract. Separate ONLINE_PREQUENTIAL training leads to frozen fresh-versus-trained evaluation. protected=false; at most 2,208 solving tasks after qualification.
- The first exact long FAST LIVE snapshot records 196/256 generation requests: 156 verified unique puzzles, 19 limits and 21 unmet generation profiles. All 156 accepted solutions were independently checked for validity and uniqueness. No solving batch appears in that snapshot.
- Earlier HF3 pilot: P1 260 SOLVED / 374 tasks; P2 72 / 90. All 332 solved grids were validated by the architect. MDL_DCC solved no more than the matched constant-predictor actuator and cost about 37.8% more WU in the main P1 comparison; seeded MRV and Luby restarts were cheaper.

Boundary: software acceptance is not causal superiority. An MDL- or DCC-specific advantage and a human learning benefit remain unestablished. R3L1 figures below retain their original version and are not results of R6.

Current arena, design and exact artifact identities:
https://www.mdlxdcc.org/crp/AI8_Sudoku_R6.html

Historical R3L1 and process evidence:

Earlier anchors:

- 1,000 solve logs
- correlation rho .85 between smarter algorithms and more compressible processes
- later Demon comparison: 6,156 operator-matched run-summary rows across 29 gates
- eight retained promotion candidates

R3L1 large-confirmation snapshot, dated 2026-09-04:

- frozen campaign: 10,000 fresh generated puzzles × 12 frozen policies = 120,000 matched rows
- first exact-byte live extract SHA3-256: ef4aa0db52043270e77747146393cc629f7c3d5cc9614d8a366877f2858f1c3c
- valid journal rows: 12,043
- complete 12-policy source blocks: 1,002
- partial blocks at extraction time: 2
- correctness, solution-match, valid-solution and given-consistency failures: 0
- completed common root-stuck cohort in the snapshot: 421

Descriptive first-snapshot totals on that root-stuck cohort:

- AI8 depth-2: 356 backtracks
- propagating MRV ascending: 851 backtracks
- provisional reduction: 58.17%
- fixed residual: 409 backtracks
- matched fixed sham: 826 backtracks
- provisional reduction: 50.48%
- fixed MDL: 445 versus fixed residual 409; MDL currently worse, with only 9/421 search-path divergences
- adaptive MDL: 378 versus adaptive residual 333; MDL currently worse, with only 12/421 search-path divergences

Interpretation:

The first snapshot supports a strong provisional whole-system and informed-lookahead signal. It does not support early MDL-specific or DCC-specific promotion. The fact that the residual controls currently outperform their MDL twins is evidence that the arena can falsify favored components rather than merely confirm a narrative.

Boundary:

This is a nonterminal first snapshot, not the frozen final result. Split B, pooled bootstrap lower bounds, wins/losses under the full contract, leave-one-out stability and final promotion gates remain uncomputed. No arbitrary-Sudoku, cross-domain, AGI, consciousness or P=NP claim follows.

Routes:
https://www.mdlxdcc.org/c/Sudoku_R3L1_Live_Snapshot_20260904.html
https://www.mdlxdcc.org/S/
https://www.mdlxdcc.org/crp/MDLxDCC-Arenas.html

## 6. P=NP-like practical behaviour

The site does not claim a proof of P=NP, a universal polynomial-time algorithm or worst-case immunity to adversarial instances.

It highlights a narrower empirical phenomenon:

The formal search space remains enormous. The effective search space can collapse.

On structured tested instances, discovering a better representation and dynamically governing search can make the useful region much smaller and more navigable than brute-force enumeration suggests. This is described as P=NP-like practical behaviour, not as a complexity-class result.

## 7. AI lifecycle applications

The same bounded-choice problem appears throughout AI:

- data sources, batches and labels
- tokenization, features and latent representations
- neural architecture search
- curriculum, objectives and training compute
- inference: continue, branch, call a tool, switch model or stop
- tool routing and verification
- multi-agent roles, communication and foreground selection
- memory promotion, quarantine, consolidation and forgetting
- research hypothesis and experiment allocation
- AI8 continuity and self-selection

The AI claim is not “another model.” It is a candidate governor for the model-development and operating lifecycle.

## 8. Domain horizon and economic scenario

The public map contains 64+ candidate application domains. Inclusion is not evidence that MDL×DCC wins in a domain. Each must earn its own baseline, oracle, matched budget, result and claim boundary.

A $3T+ modeled annual economic horizon remains on the homepage only inside a collapsed optional section. It is explicitly:

- a scenario map
- not a forecast
- not a valuation
- not achieved savings
- not evidence for the kernel
- conditional on wide adoption and domain-specific validation
- ASI excluded

Domain map: https://www.mdlxdcc.org/crp/MDLxDCC.html

## 9. How the thesis loses

L01 Simple controller: a cheaper matched scheduler reproduces the same stability, reopening and outcome.

L02 Hidden hand-coding: transfer depends on undeclared domain-specific rules instead of the shared kernel.

L03 Full cost: model, residual, decoder, tuning, communication or compute erase the apparent gain.

L04 Holdout failure: advantage disappears on frozen held-out tasks, families, scales or regimes.

L05 No replication: an external team cannot reproduce the result from code, inputs and declared protocol.

L06 Meta-cost: the recursive governor costs more than the search, communication or errors it saves.

## 10. Bounded public-first claim

Cutoff: 2026-09-03

Exact comparison target:

1. alternative-representation generation
2. full-cost MDL selection
3. dynamic stable-but-reopenable search governance
4. multi-level recursion
5. self-selection of the governor
6. measured transfer of the same architectural grammar across heterogeneous domains

Current bounded statement:

In the public systems examined so far, the targeted review found no earlier publicly documented practical system matching the complete six-part conjunction. 8Z–MDL×DCC is therefore presented as the first publicly documented practical system of this specific kind, open to correction by dated public prior art.

Limits:

- not an exhaustive patent search
- not patentability or freedom-to-operate advice
- cannot exclude private or unpublished work
- comparison concerns the complete conjunction, not the novelty of the individual ingredients

Named neighbours include DreamCoder, PowerPlay, AIXI, AlphaEvolve, ADRS and vLLM Semantic Router.

Search log: https://www.mdlxdcc.org/c/MDLxDCC_Public_Prior_Art_Log.html

## 11. Main routes

Home:
https://www.mdlxdcc.org/

Arena programme:
https://www.mdlxdcc.org/crp/MDLxDCC-Arenas.html

8zCoding executable AI work cell:
https://www.mdlxdcc.org/crp/8zCoding.html

Sudoku R6 HF3 HF1 arena and causal study:
https://www.mdlxdcc.org/crp/AI8_Sudoku_R6.html

Sudoku R3L1 first live snapshot:
https://www.mdlxdcc.org/c/Sudoku_R3L1_Live_Snapshot_20260904.html

Technical cross-domain outreach:
https://www.mdlxdcc.org/c/MDLxDCC_Outreach.html

8Z compression origin:
https://www.mdlxdcc.org/8Z/8Z_compression.html

64+ domain map:
https://www.mdlxdcc.org/crp/MDLxDCC.html

AI8 × Mind of Minds outreach:
https://www.mdlxdcc.org/c/AI8_MoM_Outreach.html

AI8 architecture:
https://www.mdlxdcc.org/AI8/

Mind of Minds deep research:
https://www.mdlxdcc.org/c/mind-of-minds.html

Human–AI invention method:
https://www.mdlxdcc.org/crp/AI8_Reasoning.html

Full work map:
https://www.mdlxdcc.org/w/

Core research papers:
https://www.mdlxdcc.org/crp/

Continuity branch:
https://www.mdlxdcc.org/c/

Consciousness and ontology branch:
https://www.mdlxdcc.org/acp/

Preserved previous dense index:
https://www.mdlxdcc.org/indexp.html

## 12. Challenge protocol

Bring one problem with too many possibilities and a trustworthy oracle.

The proposed collaboration flow is:

1. freeze task, baseline and budget;
2. build representations and governor candidates;
3. test on holdouts, ablations and strong rivals;
4. publish win, tie or loss, including causal credit, limitations and salvage.

For a continuing reviewer dialogue:

1. preserve the initial website-only verdict as a dated baseline;
2. request the exact source package or run artifact needed to test a disputed claim;
3. inspect or execute the supplied evidence under a bounded protocol;
4. state explicitly which claims, confidences, scores and verdicts stand, narrow, strengthen, revise or retract.

Allocation and validation must remain separate authorities.

Contact: Bojan Dobrečevič (BD) · bd@siol.net